Optimize MVM_serialization_read_int so as not to need memcpy calls

We only do this on CPU's that allow int64 unaligned memory accesses.
Technically the switch works on any CPU since the compiler just
optimizes it, but I think it may be faster to use memcpy in those
instances.

src/6model/serialization.c

@@ -1601,8 +1601,11 @@ MVMint64 MVM_serialization_read_int(MVMThreadContext *tc, MVMSerializationReader
         if (read_at + 8 > read_end)
             fail_deserialize(tc, reader,
                              "Read past end of serialization data buffer");
-
+#ifdef MVM_CAN_UNALIGNED_INT64
+        *((MVMuint64*)&result) = *((MVMuint64*)read_at);
+#else
         memcpy(&result, read_at, 8);
+#endif
 #ifdef MVM_BIGENDIAN
         switch_endian(&result, 8);
 #endif
@@ -1628,7 +1631,28 @@ MVMint64 MVM_serialization_read_int(MVMThreadContext *tc, MVMSerializationReader
         switch_endian(write_to, need);
     }
 #else
+#   ifdef MVM_CAN_UNALIGNED_INT64
+    /* GCC and Clang both optimize this */
+    switch (MVM_EXPECT(need, 2)) {
+        case 7:
+            ((MVMuint8*)&result)[6] = read_at[6];
+        case 6:
+            ((MVMuint8*)&result)[5] = read_at[5];
+        case 5:
+            ((MVMuint8*)&result)[4] = read_at[4];
+        case 4:
+            ((MVMuint8*)&result)[3] = read_at[3];
+        case 3:
+            ((MVMuint8*)&result)[2] = read_at[2];
+        case 2:
+            ((MVMuint8*)&result)[1] = read_at[1];
+        case 1:
+            ((MVMuint8*)&result)[0] = read_at[0];
+            break;
+    }
+#   else
     memcpy(&result, read_at, need);
+#   endif
 #endif
 
     /* Having pieced the (unsigned) value back together, sign extend it:  */